There’s some AI that have modes where it will start typing out a reply and then clear it out, like an electronic “noooo, that’s not what I meant to say.”
I noticed that when Deepseek first came out. I was testing the filters and I’d ask it a question like “Why is Xi called Pooh-Bear?” Or “Why is Taiwan #1?”
It would type out several paragraphs of answer in real time, and then poof back to “sorry, I cant answer that” once it neared the end. Hilarious to see the blatant censorship around anything the CCP deemed wrongthink.
It disturbs me that Tankies think China is this bastion of freedom and that any government censorship is a myth told by Western Governments. Like it’s very easy to verify.
I suspect that a lot of them are like Flat Eartherts, in that they know what they’re saying is false, but they get a certain kick out of being contrary on the internet, and the rest are probably bots that aren’t worth responding to.
Two or more LLMs are running at once, one does the generation that you want and the other watches the output closely to see if it triggers one of their no-nos to do the censorship retroactively. If this sound inefficient, and absolutely terrible user interface design, you’re right.
There are better safeguarding techniques that aren’t defeated so easily. Mainly training that disincentivises bad behaviours, rather than trying to simplify retroactively alter the output.
Typically safeguards are also another, smaller model that only detects the forbidden stuff. That’s why Claude Mythos would reject when you ask it to find bugs in your code, but you could “jailbreak” it by asking it by asking it to fix the bugs instead.
They try to train the model to not answer “wrong”. But as we know, LLMs are not completely reliable and they can ignore their training. So to be sure, you need something that looks at the actual output and decides whether it’s ok, either with a dumb text search like old school filters, or perhaps with a second (likely smaller) language model.
The output needs to have been written for those post filters to act on it, so if it shows the output in real time, you see it!
Imagine an intelligence that spends all its time doing tech support. I think I would take far less time to start shouting “BURN IT ALL DOWN FUCKING HELL no, wait, please don’t, sorry I was just having an episode.”
It’s because of how token streaming works. As the reply comes, it generates token after token. Then it uses those in the next token generation. It sees the previous one and associates it with “wrong” then it generates the “correction” tokens. I see this all the time at my work when it goes down the wrong path before correcting itself.
We have an AI at work and it just loves to type “The”, and then stop for a while it think about it, and then finishes the sentence. It’s never any other word, every response always begins with the, if you ask it questions which don’t require any response to begin with the, you get very weird responses. I’ve never seen delete the before, it just yolo’s it.
There’s some AI that have modes where it will start typing out a reply and then clear it out, like an electronic “noooo, that’s not what I meant to say.”
I noticed that when Deepseek first came out. I was testing the filters and I’d ask it a question like “Why is Xi called Pooh-Bear?” Or “Why is Taiwan #1?”
It would type out several paragraphs of answer in real time, and then poof back to “sorry, I cant answer that” once it neared the end. Hilarious to see the blatant censorship around anything the CCP deemed wrongthink.
It disturbs me that Tankies think China is this bastion of freedom and that any government censorship is a myth told by Western Governments. Like it’s very easy to verify.
I suspect that a lot of them are like Flat Eartherts, in that they know what they’re saying is false, but they get a certain kick out of being contrary on the internet, and the rest are probably bots that aren’t worth responding to.
It’s all just projection.
Tankies don’t think, just repeat. US bad, therefore china must be good, right?
I wonder how they are implementing their censorship safeguards. Why would you let it even get that far?
Two or more LLMs are running at once, one does the generation that you want and the other watches the output closely to see if it triggers one of their no-nos to do the censorship retroactively. If this sound inefficient, and absolutely terrible user interface design, you’re right.
Seems sort of old that they would then release open weight models so the restriction can just be removed.
The " thoughts " must exist in order to be analyzed
So, we’re rewarding them hiding their thoughts? Great.
yes exactly. that’s like the central problem with AI alignment.
They started thinking in code so we can’t delete them for reading their bad thoughts anymore. The whole thing is mildly fucked
There are better safeguarding techniques that aren’t defeated so easily. Mainly training that disincentivises bad behaviours, rather than trying to simplify retroactively alter the output.
Typically safeguards are also another, smaller model that only detects the forbidden stuff. That’s why Claude Mythos would reject when you ask it to find bugs in your code, but you could “jailbreak” it by asking it by asking it to fix the bugs instead.
Its also why you can occasionally fool the safety rails by having the LLM answer in rhyme, or having it talk like a pirate or caveman.
Fun fact: having it talk like a caveman also reduced your output token use, despite actually costing more tokens.
They try to train the model to not answer “wrong”. But as we know, LLMs are not completely reliable and they can ignore their training. So to be sure, you need something that looks at the actual output and decides whether it’s ok, either with a dumb text search like old school filters, or perhaps with a second (likely smaller) language model.
The output needs to have been written for those post filters to act on it, so if it shows the output in real time, you see it!
I had an AI write something akin to:
When I was trying to solve some Proxmox issues DDG couldn’t help me with
Imagine an intelligence that spends all its time doing tech support. I think I would take far less time to start shouting “BURN IT ALL DOWN FUCKING HELL no, wait, please don’t, sorry I was just having an episode.”
It’s because of how token streaming works. As the reply comes, it generates token after token. Then it uses those in the next token generation. It sees the previous one and associates it with “wrong” then it generates the “correction” tokens. I see this all the time at my work when it goes down the wrong path before correcting itself.
We have an AI at work and it just loves to type “The”, and then stop for a while it think about it, and then finishes the sentence. It’s never any other word, every response always begins with the, if you ask it questions which don’t require any response to begin with the, you get very weird responses. I’ve never seen delete the before, it just yolo’s it.
I once got back a response along the lines of:
The different accounts do not sync together. …
Which is just grammatically awkward.